Agent Optimization: How Mortgage Lenders and Banks Get Onto a Personal Agent's Shortlist

From Discovery to Consideration

Tristan Green of Cornerstone Advisors put the shift plainly in How banks can win in an agent-driven world: the important question is less whether consumers will let AI move their money than whether they will let it tell them where to move it. His recommendations follow from that — simplify products, pursue what he calls agent optimization, lean on strengths that do not reduce to a number, and measure how often you are considered rather than only how often you convert.

The numbers behind the argument are worth stating carefully. The piece cites an Experian study in which 82% of respondents said they trust language models to compare loans. That is trust in comparison, not in decision-making, and it is enough. In mortgage specifically, LendingTree found that 54% of homebuyers got just one offer, from one lender, on their most recent mortgage. If comparison becomes something an agent does in the background, the borrowers who would have taken one quote may get several without doing any more work.

So the job that search engine optimization did for the last twenty years — get found — has a successor: get considered. This post is about what that asks of a lender's or a bank's public information, and where the existing rules still bite.

What an Agent Reads

A person comparing mortgages reads a headline rate, a few sentences and maybe a fee table. An agent comparing on a person's behalf can read everything, and it will tend to prefer what it can parse with confidence. Four things decide whether a lender or bank is legible to it.

Products as data. Programs, terms, eligibility rules, loan limits and the scenarios each product fits, published in a structured form rather than only as marketing copy. A deposit product's tiers and conditions, likewise. An agent that cannot tell whether a borrower qualifies is unlikely to put that product on a shortlist.

Prices with their assumptions. A rate without its assumptions — credit score, loan-to-value, points, lock period, occupancy — is not comparable, and an agent that cares about accuracy may discount it or ask elsewhere. A rate published with its scenario is comparable.

Fees and conditions in one place. The fees and conditions a person discovers late are exactly what a thorough agent will look for first. Patrick Collison argued recently that personal agents could act as a "structural subsidy for product quality", because they make better products easier to tell apart. Whether or not that holds across the economy, it is a reasonable bet for loan pricing, where the differences hide in fees and terms.

What you will not do. A clear statement of what the lender does not offer — no investment properties, no bank-statement loans, no business accounts — saves the agent and the consumer a wasted application. The Personal Agent Protocol draft makes this a first-class idea: a business lists the tasks it excludes, in the customer's own words, with alternatives.

The Advertising Rules Do Not Care Who Is Reading

Publishing more pricing detail in machine-readable form does not create a new compliance regime. It runs into the existing one, and lenders should assume that a rate published for agents is an advertisement in the same way a rate on a web page is.

For closed-end credit secured by a dwelling, Regulation Z requires that an advertisement stating a rate of finance charge state it as an annual percentage rate, and allows the simple annual rate alongside it only under conditions (12 CFR 1026.24(c)). Stating certain terms — the down payment, the number of payments, the amount of any payment or of any finance charge — triggers additional disclosures (1026.24(d)). A structured rate feed that carries a rate and a payment, without the APR and the terms the rule requires alongside them, is the same violation it would be on a billboard.

For deposits, Regulation DD requires that an advertisement stating a rate of return state it as an annual percentage yield (12 CFR 1030.8(b)), with related disclosures when it does. We wrote about the recommendation side of that rule in AI deposit product recommendations.

The practical design is to publish the full disclosure set as part of the same structured record as the rate, so an agent that takes the rate takes its required context with it. That is also simply better data: an agent comparing APRs across lenders is comparing something closer to the cost of the loan than one comparing note rates.

Fair Lending Applies to the Shortlist Too

There is a quieter risk in pricing for agents. If a lender shows agents different products or prices than it shows people, or different prices to different agents, it has created a new channel whose outcomes need to be monitored like any other. A pricing difference that correlates with which agent a borrower uses, and an agent population that correlates with a protected characteristic, is a fair lending question even if nobody intended it. Our fair lending and disparate impact post covers the analysis; the simplest control is to publish one price for one scenario, regardless of who is asking.

Measure Consideration, Not Only Conversion

If the agent is the one deciding what to show, the funnel a lender can observe starts later than it used to. Cornerstone's suggestion — track how often you are on the shortlist — needs some instrumentation to be more than a slogan.

• Count agent reads. Requests for rate and product data from identified agents are the closest thing to an impression. A lender that can tell agents apart, as the protocol's signed client_id allows, can count them by agent.
• Count agent questions. Follow-up questions an agent asks about a specific scenario are a stronger signal than a page view: the lender made the first cut.
• Count handoffs. The moment an agent moves from reading to asking the borrower to sign in or apply is the moment consideration turned into intent.
• Ask why you lost. When an application started by an agent goes elsewhere, the reason is often in the comparison the agent ran — a fee, a condition, a slower answer. Agents can be asked; some will say.

What Does Not Reduce to a Number

Green's third recommendation is the one most easily forgotten in a piece about data: emphasize what an agent cannot quantify. Local knowledge, a loan officer's judgment on a complicated file, speed through a hard condition, a servicer's record with borrowers in trouble. These still matter to a borrower, and an agent that is on the borrower's side will try to represent them. The lender's job is to make them visible in a form the agent can cite — reviews, published turn times with their basis, the programs where the lender has real depth — rather than in adjectives.

The rest of what a lender needs in order to be reachable once it is on the shortlist — agent identity, borrower consent, scopes and approvals, and an audit trail — is in the personal agent readiness guide.